Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/maxiricalde/ProfeLedesma/llms.txt

Use this file to discover all available pages before exploring further.

​
Why Temporal Order Matters

For time-series problems, the train/test split must respect temporal order. Unlike tabular data where random shuffling is fine, shuffling a time series leaks future information into the training set, leading to overly optimistic evaluation. A model trained on data that includes future samples will appear to perform better than it actually does in production — because it has already “seen” the future.

​
The Workshop Strategy

The simplest valid approach used in this workshop is a year-based split: use 2020 data as training and 2021–2023 data as test. 
dfTrain = df[df.datetime.dt.year == 2020].dropna()
dfTest  = df[df.datetime.dt.year > 2020].dropna()

​
Why This Works

The entire training set is strictly earlier than the test set — there is no temporal overlap and therefore no leakage possible. Every prediction the model makes during evaluation is a genuine out-of-sample forecast.

​
Alternative Strategies

The simple year-split above is sufficient for this workshop. More rigorous approaches exist for production systems:
  • Rolling window validation — a fixed-size window slides forward in time; each window position defines one train/test split.
  • Walk-forward validation — the model is retrained at each step as new data arrives, simulating real deployment.
  • Expanding window — the training set grows with each step while the test set always starts immediately after the current training endpoint.
These methods produce multiple evaluation scores and give a better picture of how the model generalises across different time periods.

​
Inspecting the Splits

After splitting, use .describe() on both sets to confirm they have comparable GHI distributions. Very different statistics (e.g. a much higher mean in the test set) would signal seasonal imbalance that could skew metric interpretation. 
dfTrain.describe()
dfTest.describe()
Never use train_test_split(shuffle=True) from scikit-learn on time-series data. Always split by time index. Shuffling randomly mixes past and future samples, making the evaluation meaningless.
The dropna() call ensures only rows with valid (non-NaN) GHI readings are included. NaN values were inserted earlier by QC filters and the resampling threshold step — rows where too few raw measurements contributed to an hourly average are marked as missing and excluded here.

Build docs developers (and LLMs) love

Get started for freeTalk to us